Download Fulltext - Brunel University Research Archive

Unconstrained evolution of analogue computational
“QR” circuit with oscillating length representation
Yerbol Sapargaliyev1, Tatiana G. Kalganova 1
1
School of Engineering and Design, Brunel University
Uxbridge, Middlesex, UB8 3PH, UK
{Yerbol.Sapar, Tatiana.Kalganova}@brunel.ac.uk
Abstract. The unconstrained evolution has already been applied in the past
towards the design of digital circuits, and extraordinary results have been
obtained, including generation of circuits with smaller number of electronic
components. In this paper unconstrained evolution, blended with oscillating
length genotype sweeping strategy, is applied towards the design of "QR"
analogue circuit on the example of circuit that performs the cube root function.
The promising results are obtained. The new algorithm has produced the
excellent result in terms of quality of the circuit evolved and evolutionary
resources required. It differs from previous ones by its simplicity and represents
one of the first attempts to apply Evolutionary Strategy towards the analogue
circuit design. The obtained result is compared with previous designs.
Keywords: Evolutionary hardware design, Evolutionary circuit diagnostics and
testing.
1 Introduction
The Evolvable Hardware (EHW) is one of the most promising areas of today’s
electronics. Evolutionary Algorithm (EA) applied towards reconfigurable hardware
enables to find a solution among global solution space. The EHW where the ultimate
goal is a circuit design is called Evolutionary Electronics [1], [4]. The evolutionary
electronics gives the alluring opportunity for an amateur in the field of Electronics to
reach out the same results as professional, possessing mostly the knowledge of
Darwinian’s laws and the inspiration. The EA, navigated by a fitness values, provides
randomly created and mutated chromosomes. Each chromosome encodes the structure
for a circuit and has to be evaluated a fitness function assigning a fitness value. The
fitness value shows how close the current hardware structure by its behavioral
characteristics to the required one. The circuits evolved may have unconventional
designs and less of all depend on personal knowledge of a designer.
For instance, using simulation software (extrinsic EHW), low-pass filters [2]-[9],
high-pass filters [10] and amplifiers [3], [4], [5], [11] are successfully designed with
the help of EA. Moreover, the structure and element parameters of reconfigurable
hardware by itself could be set as an evolutionary target [13]. To evolve the base to be
evolved for further evolutionary designs seems quite promising and interesting
perspective, but implies the strict requirements towards the evolutionary technique.
The evolutionary technique is a set of rules according to which sweeping strategies
[4], the circuit growth strategy, parameters of EA and the circuit representation
technique are managed.
In [18] the unconstrained evolution, both spatially and temporally, intrinsically has
been applied towards the digital reconfigurable hardware -FPGA. By releasing the
full repertoire of behaviors that FPGA can manifest, namely, allowing any
connections among modules, letting the evolution to evolve the granularity of
modules as well as the regimes of synchronization, evolution has been able to find a
highly efficient electronic structure, which requires 1-2 orders less silicon area to
achieve the same performance as conventional design does. Ones fully unconstraining
the design methodology rules, the natural behavior of analogue elements started to be
exploited inside a circuit.
In analogy to this approach, the absolutely unconstrained evolution is applied in
[19] towards the originally analogue circuits. In this sense, the range of circuitstructure-checking rules at the netlist composition stage, prohibiting the invalid circuit
graphs, are regarded as the main constraints for the design methodology.
In this paper, we utilize the same algorithm as in [2] trying to evolve the analogue
circuit that performs the cube root function. Except [2], for this task no circuit was
found in the published literature.
The framework of work is to demonstrate the capability of evolutionary process to
evolve complex non-linear computational circuits with limited amount of
computational resources used, not to apply the developed approach to real world
applications. The accuracy of obtained results can be improved if the number of
points considered during evaluation process has been increased.
Absolutely agreeing with the factors set in [5] as favorite for evolutionary circuit
design: E-12 series of component parameters and modesty in element amount within a
circuit, we also set as important the simplicity of the evolutionary technique. Last one
makes the experiments to be reconstructed and the same results easier to retain. We
propose the simplest oscillating length genotype (OLG) sweeping strategy [3] that
together with unconstrained evolution gives excellent results. The evolutionary
technique as well as the obtained results is compared with ones published previously
on evolution of analogue filters.
The next section overviews the previous work in the area. Section III introduces
the whole evolutionary technique. Section IV describes the experimental results
together with comparison between the results. And, finally, the last section concludes
the paper.
2 Previous Work
The importance of analogue evolutionary circuit design is well described in [3]. In
Table 1 we included most of the works in evolutionary analogue circuit design. Most
of the works start from evolving a passive low-pass filter. Last one is a convenient
tool for probation of evolutionary technique and tuning the EA parameters towards
the more sophisticated designs [2], [3], [13]. Our technique also was probed on this
task [19] and excellent results were retained.
The considerable results were obtained in [2]. They used Genetic Programming
(GP) circuit-constructing program trees approach with four kinds of circuitconstructing functions. They also used automatically defined functions and potentially
enabled certain substructures to be reused. They got three kinds of computational
circuits including cube root function with very precise performance. The main
drawback of this experience is the large computing power required as well as the
complicacy of the methodology.
Works [4] and [6] gave the comparison between GP and Genetic Algorithm. The
first work was made as analogy to biology concept with comparison of different types
of variable length chromosome strategies, while in the second one there was intrinsic
evolution of real hardware for robustness purposes.
According to [4], sweeping strategies refer to the way in which the different
dimensionalities of the genome space are sampled by the EA. There are three kinds of
sweeping strategies introduced by Zebulum et al in [4]: Increasing Length Genotypes
(ILG), Oscillating Length Genotypes (OLG) and Uniformly Distributed Initial
Population (UDIP). The OLG strategy is a variation of the ILG strategy in which the
genotypes are also allowed to decrease in size. The main purpose of this strategy is to
create pathways from large to smaller genotypes with similar fitness values. In UDIP
instead of starting with a population of small genotypes, the initial genotype sizes are
now randomly assigned to values ranging from one to the maximum number of genes.
In [4] all tree strategies were applied to evolving the LCR (filters) and QR
(amplifiers) circuits, and ILG and OLG strategies have shown superior results.
Table 1. Recent advanced on the evolution of analogue circuits, SA is Simulated Annealing
Koza et al [2]
Lohn, Colombano [3]
Goh, Li [5]
Zebulum, et al [4]
Grimbleby [7]
Dastidar, et al [11]
Ando, Iba [6]
Sripramong et al [12]
LCR circuit [19]
Sapargaliyev,
QR circuit
Kalganova
(proposed)
Year
1997
2000
2000
1998
1999
Type of
EA
GP
GA
GA
GP,GA
GA
Circuitstructurechecking
rules
Partially
Yes
Yes
Yes
Data n/a
2005
2003
2002
2006
GA
GP,GA
ES+SA
ES
Yes
Yes
Yes
No
GA
GP,GA
hill-climbing
No
OLG
Data n/a
Fixed
OLG
2008
ES
No
No
OLG
Parameter
Sweeping
optimization strategy
No
ILG
No
ILG
No
ILG
No
ILG,OLG,UDIP
numerical
ILG
As could be noticed from the Table 1, the following gaps do still exist, which we
try to fill with work presented:
• The most of previous research in analogue circuit design used the circuitstructure-checking rules for avoiding invalid circuit graphs.
•
The most of works in low-pass filter design used the ILG strategy as a sweeping
strategy; however in work [4] the OLG has shown excellent results for analogue
circuit design, and the best for low-pass filter design.
In the frame of this work we will try to fill these gaps in someway.
3 Unconstrained evolution of “QR” circuits
Representation. We use for circuit building only tree types of elements: Qn which is
n-p-n bipolar transistor, Qp which is p-n-p bipolar transistor and R – resistors. The
linear circuit representation is proposed for use, similar to one that exploited in [4],
that is every element of a circuit is represented as a particular gene, and each gene
consists of 4 loci corresponding to element’s features: element’s name, node numbers
to each pin and parameter (only for R).
Resistor encoding:
Bipolar transistor encoding:
Re N1 N2 Pa
Qx N1 N2 N3
Fig. 1. A gin coding a resistor (left) and a bipolar transistor (right): Re-loci is the resistor’s
name; Qx-loci is the transistor’s name; N1, N2, N3 -loci are the nodes for the first, the second
and the third pins; Pa-loci is the element’s parameter. Amount of variations is shown over each
loci.
On Fig.1 (left) is a view of a gene coding a resistor, Fig.1 (right) is representation
for a bipolar transistor. The gene looks exactly the same as an element line in the
PSPICE netlist, so, there is no necessity to convert a genotype into a netlist.
For resistor Pa-loci we set 64 possible values, according to E-12 order, which
means there are 12 parameters per decade. That is we covered 5 decades plus 4 “justin-case” parameters located in upper and lower neighboring decades.
Unconstrained evolution of “QR” circuit. In [19] we called absolutely
unconstrained evolution of an analogue circuit the process of circuit netlist generation
during which no circuit-structure-checking rules applied and all the circuits are
counted as valid graphs except ones that have elements with dangling nodes and with
isolated sub-circuits. Note that the term “unconstrained” is applied to the duration of
evolutionary process, i.e. the way of obtaining the target circuit that differs from the
way how the final valid circuit structure is obtained. In other words, the automatic
design of logic circuit contains of two main stages: 1) unconstrained evolution of
logic circuit; 2) generation of final valid circuit structure.
There are two main kinds of invalidities that unacceptable for most of simulation
software and that just prohibited by most of researchers in the area (Table 1): the
nodes without DC path to ground (tackled in [2]) and loops involved inductors and/or
voltage source (tackled in [19]). In [2] and [19] these problems were tackled by using
R-support that made almost any randomly generated circuit valid.
In this work where we do not use the reactive elements (L and C), the task of
unconstraining the evolutionary search becomes easier, because both kinds of
elements that we use have internal resistances and both of them being anywhere
inside the circuit have DC paths to the ground.
However, in most of the works listed in Table 1, rules prohibiting some transistor
connections, such like emitter-to-collector connection, were applied to reduce the
solution space. These connections do not certainly produce non-convergence errors.
Emitter-collector connection leads only to temporarily tripping of that particular
transistor from the whole circuit, until the further mutation process suddenly activates
it. In this work we do not prohibit such kind of connections removing these last
constrains. Thus, since we do not have any other kinds of constrains that have to be
released, Our technique can be called as an absolutely unconstrained evolution for
“QR” circuits.
Experiment settings. The embryo circuit is the element or the number of elements
(including the voltage source), that can be predetermined for the particular circuit to
ease the further circuit growth. We define the embryo circuit correspondent to the
most popular case where the circuit is driven by a DC voltage source, source
resistance Rsource=1kΩ and the load resistance Rload=1kΩ. These three elements
compose the embryonic circuit and absolutely identical to that ones in most of the
works (Fig. 2). The output voltage is measured on the pins of Rload.
Node1
Evolving circuit
Rload
Rsource
+
-
Node2
Vsource
Node0
Fig. 2. Embryo Circuit
1. Start
Output
Voltage
Embryo chromosomes
Fitness
values
4. Fitness
Evaluation
2. ES
Population of 30,000
chromosomes
AC analysis results
Chromosome
netlists
3. Orcad
Pspice
PSPI
CE3.eThe flowchart of the experiment
Fig.
Fig.3 generally shows the algorithm of the experiment. It consists of 4 main blocks.
The PC program written in C programming language describes all 4 parts and unites
them in one code.
The Start-block provides population of chromosomes in form of PSPICE netlists.
This block includes all the data necessary for embryo circuit production: initial
parameters, element description and analysis options. Being delivered to ES block,
every chromosome at this stage is grown up from the embryo to the individuals with
the same number of genes (elements). We set this number at three.
ES part sets the particular parameters of ES, such as: mutation rate, population
size, selection criteria and termination terms. It modifies the genotype and produces
the population of chromosomes in form of cir-batch-file towards OrCAD PSPICE.
Last one utilized in non-interactive batch simulation mode.
Block 3 downloads cir-file to the PSPICE, receives the result in form of out-file
and passes it for evaluation to Block 4.
Block 4 defines the best chromosome, selects 10% fittest individuals and sends
them to the Block 1. At this stage Start block, depending on the best fitness, decides
which chromosome to send to which mutation.
As can be seen on Fig.5, the whole process consists of three kinds of operations
over the each chromosome. Depending on correlations among lengths and fitnesses of
the best and current chromosomes, each chromosome it put under the particular
mutation. Add_new_element_mutation (ANEM) is a procedure, when one randomly
generated gene is added to each chromosome except the chromosome with the best
fitness value. We set the maximum difference in length between the shortest and the
longest chromosome up to 5 genes. ANEM is applied in cases when the current
individual is longer than the best one.
To restrain the difference between the shortest and the longest chromosomes the
Delete_element_mutation (DEM) is devoted, that deletes one gene if the difference
exceeds the limit. Thus, the evolution can focus on processing chromosomes of five
different neighboring sizes.
The “circuit structure mutation” (CSM) performs mutation over any of four loci of
randomly chosen gene (Fig.1). If the mutation comes to a pin connection, the whole
structure of a circuit is changed. Despite the total amount of elements stays
unchangeable, the number of nodes of a circuit could be reduced or increased. CSM
also performs the parameter optimization procedure.
Fitness Function and Termination Criteria. The target for evolutionary search is to
evolve an analog circuit which output voltage is the cube root of its input voltage. To
enable ourselves to make the estimation the final results from experiment we have set
the same fitness terms as in [2]. That is, we made the PSPICE simulator to perform a
DC sweep analysis at 21 equidistant voltages between –250 mV and +250 mV for the
cube root. Fitness value is set to the sum, over these 21 fitness cases of the absolute
weighted deviation between the target value and the actual output value voltage
produced by the circuit. The smaller the fitness value, the closer the circuit to the
target. We set for fitness to penalize the output voltage by 10 if it is not within 1% of
the target voltage value. The error circuits, that mostly are non-convergent, are not
analyzed by simulator and assigned to the worst fitness value that never could be
reached by other circuits.
We set as a termination criteria reaching either of the following conditions: the
fitness value does not improve over 20 generations (600 000 individuals), or the best
circuit reaches more than 100 elements, or the best fitness value reaches 0.5.
According to experience, the most frequent termination way was the stop in fitness
increase.
4 Experimental Results
The result presented is the best out of 20 runs at 10 different PCs with different seed
for the random number generator. The Evolutionary Strategy with linear
representation and oscillating length genotype is utilized. The total population was
30 000 individuals, mutation rate 5%.
At generation 3 the best individual (No 24999) with 3 genes (in addition to embryo
elements) shown the fitness 65.57. The circuit that this chromosome describes is
presented on Fig. 4 and has 2 transistors and one resistor.
The next notable result appeared at generation 15 (No 23882) with 14 genes (in
addition to embryo elements), which describes a circuit with 7 transistors, 1 diode
(transistor whose collector is connected to the base) and 6 resistors. This circuit,
pictured on Fig.5, has the fitness 5.53.
And, finally, the circuit that evolution reached after 133 generations (No 24318),
and which evolution was not able to improve during the next 20 generations, appeared
with totally 38 elements (in addition to embryo elements): 24 transistors, 12 resistors
and 2 diodes (Fig. 6). The fitness of this circuit achieved 2.27.
The average error-circuits per population was 4-5%, mostly they are nonconvergent.
Fig. 5. Best circuit from generation 15
Fig. 4. Best circuit from generation 3
Table 2. Comparison of evolved cube root computational circuits
1-st circuit
Element
Fitness
number_1 value_1
filter,
3
77.7
Evolved
Koza, 1997 [2]
Evolved filter in this
work
Gain
2-nd circuit
Element
Fitness
number_2
value_2
18
26.7
3-rd circuit
Element
Fitness
number_3 value_3
50
1.68
3
65.6
14
5.53
38
2.27
0
15.57%
22.22%
79.29%
24.00%
-35.12%
The cube root computational circuits designed either conventionally or using
evolutionary process is very rare to find in literature. Up to date we were unable to
identify the work that will design non-linear computational circuits such as cube root,
etc. conventionally [2]. To decide on efficiency of the proposed evolutionary
technique we have only one work directly to address to [2]. Since we set the same
fitness function, we can directly compare the circuits and their corresponding fitness
values. The result of comparison is presented in Table 2.
5 Conclusion
In this paper we applied the unconstrained evolution towards the analogue circuit
design on the example of “QR” computational circuit performing the cube root
function.
The method utilized here is much easier than that one applied in [2]. While last
approach, with help of reusable sub-constructions, successfully evolved circuits with
large amount of elements, our method, as it could be seen from Table 2, succeeds in
small and middle sized circuits.
Fig. 6. Best circuit from generation 113
As it could be seen from Table 3, the computer resources, and thus potentially
time, in our attempt much lower, reaching in average 90%.
The proposed method has shown its potential for further improvements by getting
the fitness that close to one in [2]. If compare the final results, the shortage in fitness
value is almost the same (35%) as the gain in element economy (24%).
Table 3. Comparison of computational resources
1-st circuit
Evolved filter, Koza, 1997 [2].
640,000
individuals
per
population
Evolved filter in this work.
30,000
individuals
per
population
Gain
Gen. #0,
0-640,000
2-nd circuit
3-rd circuit
Generation #3
90,000
Gen. #17,
10,240,00010,880,000
Generation #15
443,882
Gen.#60
37,760,00038,000,000
Generation #133
3,984,318
Up to 86%
Up to 96%
Up to 89%
This paper has shown one of the first successful attempts of application of
Evolutionary Strategy towards the analogue circuit design.
The oscillating length genotype sweeping strategy with capability of evolution to
focus on the limited genotype length dispersion proved its high surviving character.
Despite our attempt to succeed the works [2], last one, accomplished by the leading
research group in evolvable hardware, exemplifies for us the perfect result ever made
by evolutionary tool on the analogue circuit design.
Note that the GA convergence and averaged result over 20 or 100 runs have not
been completed since there is not existing work these parameters can be compared
with. This is subject for future work.
The analysis of the evolved circuit structures demonstrated in this work as well as
in [2], shows that in order to provide the circuit structures stable to variations in
temperature, supply voltage, noise, etc. it is necessary to complete the second
evolutionary process that will concentrate on evolution of stable wiring. This will
make the first step towards evolution of real-world analogue circuits, stable to
changes in the environment, and, consequently, the evolvable hardware design
approach can reach the competiveness stage with conventional design. This is also the
subject for future work.
Acknowledgement
This work has been carried out with support from Concept Engineering Company
who made a donation for this research.
References
[1]
[2]
Thompson, A.: Hardware Evolution: Automatic Design of Electronic Circuits in
Reconfigurable Hardware by Artificial Evolution. D.Phil. thesis, University of Sussex,
Brighton, Sussex, England (1996)
Koza J.R., Bennett III F.H., Forrest H, Lohn J, Dunlap F., Andre D., Keane M.A.:
Automated synthesis of computational circuits using genetic programming. In: 1997
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
IEEE Conference on Evolutionary Computation, pp. 447–452 Piscataway, NJ: IEEE
Press (1997)
Lohn J.D., Colombano S.P.: Automated Analog Circuit Synthesis using a Linear
Representation. In: The 2nd International Conferene on Evolvable Systems: From
Biology to Hardware, pp.125- 133 Springer-Verlag, Berlin (1998)
Zebulum R.S., Pacheco M.A., Vellasco M.: Comparison of different evolutionary
methodologies Applied to electronic filter design. In: IEEE Conf. on Evolutionary
Computation, pp. 434-439 Piscataway, NJ: IEEE Press (1998)
Goh C., Li Y.: GA automated design and synthesis of analog circuits with practical
constraints. In: The Congress on Evolutionary Computation, Vol. 1, pp. 170-177 (2001)
Ando S., Iba H.: Analog Circuit Design with a Variable Length Chromosome. In:
Congress on Evolutionary Computation, pp.994-100 IEEE Press (2000)
Grimbleby J.B.: Hybrid genetic algorithms for analogue network synthesis. In: Congress
on Evolutionary Computing (CEC99) pp. 1781-1787 Washington USA (1999)
Fan Z., Hu J., Seo K., Goodman E., Rosenberg R., Zhang B.: Bond Graph Representation
and GP for Automated Analog Filter Design. In: E. Goodman (ed.) 2001 Genetic and
Evolutionary Computation Conference Late-Breaking Papers, pp.81-86 ISGEC Press,
San Francisco (2001)
Wang F., Li Y., Li L., Li K.: Automated analog circuit design using two-layer genetic
programming. In: Int. J. on Applied Mathematics and Computation, Special Issue on
Intelligent Computing Theory and Methodology Vol. 185, Issue 2, pp. 1087-1097 (2007)
Hu J., Zhong X., Goodman E.: Open-ended robust design of analog filters using genetic
programming. In: Genetic & Evolutionary Computation Conference (GECCO) pp.16191626 ACM New York, NY, USA (2005)
Dastidar, T.R., Chakrabarti, P.P., Ray, P.: A synthesis system for analog circuits based on
evolutionary search and topological reuse. In: IEEE Trans. on Evolutionary Computation,
Vol. 9, Issue 2, pp. 211 – 224 (2005)
Sripramong, T., Toumazou, C.: The invention of CMOS amplifiers using genetic
programming and current-flow analysis. In: IEEE Trans. on Computer-Aided Design of
Integrated Circuits and Systems, Vol. 21, Issue 11, pp.1237 – 1252 (2002)
Zebulum R., Stoica A., Keymeulen D.: Experiments on the Evolution of Digital to
Analog Converters. In: IEEE Aerospace Conference, Big Sky, Montana, USA Manhattan
Beach, CA ISBN: 0-78-3-6600-X. (Published in CD) (2001)
Hu J., Zhong X., Goodman E.: Open-ended Robust Design of Analog Filters Using
Genetic Programming. In: Genetic & Evolutionary Computation Conference (GECCO)
Vol. 2, pp.1619-1626 ACM Press, Washington, DC (2005)
Kuo T. and Hwang S.-H.: Using disruptive selection to maintain diversity in genetic
algorithms. In: Appl. Intel. 7, pp.257–267 (1997)
Brameier, M.: On Linear Genetic Programming. In: PhD thesis, University of Dortmund,
Dortmund, Germany (2004)
Vesselin K., Miller J.: The advantages of landscape neutrality in digital circuit evolution.
In: International Conference on Evolvable Systems (ICES), Lecture Notes in Computer
Science, pp. 252-263 Springer (2000).
Thompson A.: Artificial evolution in the physical world. In: Gomi (Ed.) Evolutionary
Robotics, AAI Books (1997)
Sapargaliyev Y., Kalganova T.G.: On Comparison of Constrained and Unconstrained
Evolutions in Analogue Electronics on the Example of "LC" Low-Pass Filters. In: IEICE
transactions on Electronics Vol.E89-C No.12 pp.1920-1927 (2006)